Leveraging Multiple Sources in Automatic African American English Dialect Detection for Adults and Children

This work presents a novel system which utilizes acoustic, phonological, morphosyntactic, and prosodic information for binary automatic dialect detection of African American English. We train this system utilizing out-of-domain adult speech data and then evaluate on both children's and adults' speech with unmatched training and testing scenarios. The proposed system combines novel and state-of-the-art architectures, including a multi-modal masked language model pre-trained on Twitter text data and fine-tuned on ASR transcripts as well as an LSTM trained on self-supervised learning representations, in order to learn a comprehensive view of dialect. We show robust performance for our system across age and recording conditions.